Summary
2023
Session Number:C2L-1
Session:
Number:C2L-13
Investigation of Relationships Within the Latent Variable Space of Sentence-Bert Sentence Vectors
Izumi Masato, Jin'no Kenya,
pp.451-454
Publication Date:2023-09-21
Online ISSN:2188-5079
Summary:
This paper investigates the relationship be- tween colors within the latent variable space using an image generation model that outputs images corresponding to the sentence ”color + of + car name”. In particular, we explore the relationship between the colors of different car models using the model, and find that the positions of col- ors within the latent variable space are similar among the car models, forming shapes resembling equilateral triangles. We investigate the position of yellow within the color space, finding that it may be located on a sphere formed by the colors RGB and CMY, with these colors forming triangles while moving towards each other. Overall, we offer insights into the relationship between colors within the la- tent variable space, shedding light on the structure of color representation in natural language processing models.